Research on Deep Learning-Based Method for Bearing Fault Diagnosis in TENG Under Wear Conditions

The Triboelectric Nanogenerator (TENG), as an emerging self-powered sensor, is widely used in the field of rotating machinery bearing fault diagnosis. Due to its working principle based on frictional electrification and electrostatic induction effects, the surface morphology and charge transfer efficiency of the friction layer have a significant impact on the output performance of TENGs. Under long-term mechanical motion, the friction layer may experience wear, and continuous wear can lead to surface morphology damage and even damage to the friction layer structure, gradually destroying the TENG's signal acquisition and output capabilities, causing signal degradation and bearing fault feature deviation, which results in a decrease in bearing fault diagnosis accuracy. Traditional solutions focus on material properties and structure. This article derives the mechanism of the influence of friction layer thickness on the output signal through the TENG output voltage formula and simulates different degrees of wear with friction layers of different thicknesses to conduct deep learning-based bearing fault diagnosis experiments. The experimental results show that although the CNN model can recognize TENG signals well for bearing fault classification, the bearing fault features of the worn signals shift, and the accuracy of CNN diagnosis decreases. The introduction of a one-dimensional self-attention-enhanced convolutional neural network model and an incremental learning method improved the accuracy of bearing fault diagnosis after wear and tear. This study provides theoretical support and practical solutions for long-term, stable bearing fault diagnosis in TENG under wear conditions.